Qhantom
← All articles
An apple resting on a window sill in violet light

Article

The Apple Behind the Door

Qhantom Labs • Perspective and AI omniscience

A cook leaves an apple on the kitchen table and steps out into the hall. The door closes, and inside, someone silently moves the apple across the room to a window sill.

Asking the cook where the apple is, he replies, “Last I saw, it was on the table” would be the honest answer based on what the cook knew. “It’s now across the room” would be the correct answer. But in this example the cook would not know the apple has been moved.

One world and different points of view, this is how persistent worlds should work.

As we research and aspire to make great innovations in this field, and as we build Qonvergence, we believe participants in the world should have only the knowledge they can observe themselves. What happens in the world matters after you leave.

If someone moves the apple, it should still be across the room when somebody who was not there returns, and the cook in this example, should not have omniscient knowledge that it moved when he was gone. Since the cook never saw it move, the cook should remember the apple is on the table. A restart or a new query should not undo the apple’s movement nor should it teach the cook something they never learned. This is extremely important in both games existing inside persistent worlds and in using these technologies for experiments, simulation and research.

Some games already make an attempt at these distinctions. There’s a version of the apple problem in a 2024 study by Zhou and colleagues. They simulated two strangers at a party trying to figure out whether they had a friend in common. When one model played both characters and could see both of their private friend lists, the mutual friend’s name came up much earlier than when each agent could see only its own list. The conversation looked like a discovery, but the model already had the answer and was role playing ignorance. This is a useful starting position for a question we are asking during testing in persistent worlds • what information actually reaches each person or agent and what happens to the information when time passes or a restart occurs in the world?

When a world tells an agent too much information we call this, “accidental omniscience”, when a system gives an agent knowledge on a fact it shouldn’t have access to. The cook can guess where the apple has gone, maybe a witness to the event informs the cook, but if the cook is given information to pick up the same apple from the window sill while the cook is away from the room, the world has now leaked information to the cook that he should not have. Retrieved memory or tool responses are other ways this can occur.

In a world, the effects of this accidental omniscience could be large. During an experiment a control group agent could become privy to the fact they are the control group, sticking a wrench in the test, in a game a trade a character made in private with a rival faction could become known to the people they are betraying, or a discovery that someone has yet to share can be leaked into public.

One way in which we are testing it. Run the same actions twice.

  • Version 1move the apple
  • Version 2leave the apple alone

Before the door is reopened, compare everything the cook receives • scene description, memories, available actions, tool replies. The hidden move of the apple should not change those inputs. Matching the inputs alone is not enough • both runs could theoretically be handing the cook the same private inventory, so checking the inputs against a separate, static list of what the cook is allowed to know is important.

Then restart the room and check again. Open the door and make sure the cook can see the correct location in each version. A system that hides truths is also a fail.

Making sure the world does not give away answers and information is just as important as whether an agent is smart enough to be able to discover where the apple has gone.